It seems we have some choices to make:

1. try and prevent the RAM fragmentation from happening in the first place.

2. clean up fragmentation when RAM is freed.

If we go the second route (and that's where I bet we'll go), then we 
have two choices for that:

1. coalesce RAM when it is freed.

2. coalesce RAM with a kernel thread during idle time, sometime after 
it is freed.

The first probably means excess overhead, but the second will be 
unhappy under load (i.e. if there is no idle time available, or when 
there is a heavy demand for contiguous RAM), so we should probably do 
#2 "normally" and #1 under load, so that we can be adaptive.

What have I missed?

	Erik <fair@netbsd.org>